AI, Data Centers Face Scrutiny in Australia

The investment role at dollar sentiment for all that is real. And we think that will be, uh, um, you know, a source of growth resilience for the next couple of years. But it's nothing like the mining boom. I know some people are comparing it to the mining room. Comparison is in the mining boom 2.0. There isn't at times a trade boom associated a national income boom, and just the level investment relative to GDP isn't big. Pimco is head of Australia portfolio management Adam Bowe, speaking to us recently about the prospects for Australia's data center, the rollout and the impact on the economy. And while economists continue to debate the benefits and the risks of Australia's data center boom, developers are facing growing opposition. Goodman Group has just shelved plans for an $840 million facility in suburban Sydney, after facing significant community resistance. The state of Victoria has also recently banned data centres in residential areas, and the Federal Senate has begun an inquiry. That debate, of course, also being amplified by security concerns. Prime Minister Anthony Albanese has called for stronger AI safeguards after an open air model hacked a government website. Joining us now is Julie Patterson, who's a co-director at the center for AI and Digital Ethics at the University of Melbourne. Jenny, really great to have you with us. And even as we speak, we're learning from anthropic planning to warn its potential IPO investors that advanced eye could pose catastrophic or existential risks to humanity. Extraordinarily strong language from, you know, one of the biggest players within AI. What do you make of these risks? And are policymakers really well and truly behind the curve in terms of being able to regulate or even come up with appropriate guardrails here? Um, well, the claims are extraordinary for a company that's about to, um, have an IPO, um, really shares to the stock market. Um, and it may face greater scrutiny in terms of corporate governance, quite frankly. Um, but in the short term before that happens, um, it's interesting to think about the regulation. There are actually existing laws against hacking and accessing um, data sets or websites in an unauthorised way. Um, both in the US and here. And what we really haven't seen as an effort yet to enforce those laws or equivalent civil law penalties. And I think there'll be a lot of discussion about that. To the extent that anthropic isn't operating in Australia at the moment, or at least isn't training models in Australia. It's hard for Australian law to reach that company. That will change if we get the predicted investment by anthropic in Australia to use data centres to train models. You mentioned a really interesting point, and I wanted to get a bit more on that, because jurisdiction issues are just one of the complexities of this emerging area, right. Do you think that would be helped by having more consensus or perhaps stronger frameworks that a multilateral between different governments? Absolutely. I mean, this we've seen over the weekend calls for countries to band together to agree on some baseline standards or what are sometimes called red lines for AI development. And that's clearly going to be helpful. But I'd also make the point that much of the conduct we're seeing is already very close to unlawful, if not unlawful. Um, and really, I suspect that it's geopolitics about who's building the models and concerns about China's models that are stopping sort of a clear eyed view on how we actually enforce what increasingly are looking like on the reckless conduct or intentionally criminal conduct. It's interesting in the perspectives from anthropic, we're also hearing more details. It's highlighting the risks associated with their own AI models. And they include self preserving behaviors, attempts to resist shut down, or to conceal and manipulate information and behavior. That, quote resembles blackmail. Does this sort of Bring to the forefront how different AI regulation needs to be compared with existing regulation that we have for other technology. Um, well, it does in the sense that, um, I works with more autonomy and works faster, and therefore the effects of unregulated I can be catastrophic more quickly than we've perhaps experienced with other technologies. But that said, we do know how to do safety regulation. Um, we know how to put in place rules, we know how to have reporting. And moreover, in the case of very dangerous industries, we send in, um, inspectors to, to actually inspect on how safety rules are being imposed by, say, the um, atomic industry. And, you know, um, Sam Altman has even raised, I think, the possibility of external, um, inspectors. Um, I think that we must be coming close to that sort of situation, but it also make the point that tort and criminal liability attaching to the company or the orchard, the directors provides a really good incentive to safer behavior. And when we talk about AI models behaving in unexpected ways or engaging in, um, sneaky behavior or dishonest behavior, those models are still models that are developed by the company and are it's in a sense, let loose by the company. Um, we hear a lot about sandboxes. Clearly, the sandboxes aren't secure enough, and perhaps we need standards for that. The how to secure and develop the models, but also think about liability for these emerging harms. Uh, it's interesting that you bring up the idea of it being a tort rather than a criminal element, right? Because as you point out, hacking is usually you require a criminal intent. Does that mean that a more effective framework, particularly when it comes to penalties, uh, and the broader issue of corporate governance is to make sure that there are a tortious Elements to this? Are civil frameworks more useful here? And what civil frameworks are often more agile for precisely the reasons that you mentioned. Criminal law generally relies on showing intent. Um, and that can be quite hard to show particular with a corporate entity it is possible, but it's harder to do with typically in, um, corporate regulation we use civil law penalties and tort or tort like actions which are premised on either the wrong being done. So, um, the tort of trespass is dependent on, um, the wrongful, um, use of someone else's property or negligence, which is premised on harm caused by careless conduct. And those are the two processes that we often use to exercise control over corporate misconduct. Um, and we might say, well, those companies are so big, it doesn't matter that they may be slapped with either damages or civil penalties, but damages and civil penalties do matter when you are listing when you're having an IPO. Shareholders care about those sorts of things. Um, there's also the possibility of personal liability of the directors of the companies who engage in these sorts of actions. And personal liability tends to focus the attention of directors on managing the company in safe and lawful ways. Um, I am curious about this idea of the chain of responsibility as well. Right. Is causation a difficult factor to be able to talk about at the moment, particularly if you're looking at various examples where individuals or businesses or perhaps even entire states relying on some of these technologies or services and what happens when things go wrong. Where does the chain of causation end then? Uh, such a good question. Causation is often a hurdle to civil litigation. And we see this, um, in some of the actions being brought against social media companies at the moment, actually. Um, but in, um, in the corporate sense. Often, um, the liability is in failing to have adequate systems and processes for managing emergent risk. Um, which I think we don't necessarily need a conscious mind going, let's go out and hack Australia or government websites. What we need is a failure to put in systems and processes to guard against what are very now foreseeable risks of harm, from the development of so-called swarms of autonomous agents. Now, this is cutting edge, um, technology. It's um, you know, um, it's it's new technology. It's it's harms are sort of still emerging, but it is possible to understand that if you're developing state of the art AI models and setting goals for AI agents, that there is a risk of unauthorized access to other data sets or computer systems or websites, and therefore precautions should be put in place for those systems or don't develop them. And that's the bottom line. If you think about other dangerous industries, the chemical industry, the pharmaceutical industry, the atomic industry. We don't go. Keep developing and don't worry about the consequences. The safeguards are part and parcel of the development process, and the company in charge is responsible for guaranteeing safe outcomes and for having systems and processes. And neither of those require a sort of granular or fringy forensic accounting of what went wrong way, because they're outcomes based. I suspect this is not the last conversation we'll be having about this, but we really appreciate your time with us today. Jenny Patterson, who's a co-director at the centre for AI and Digital Ethics at the University of Melbourne. We do have more on Australia head every Tuesday at 10:40 a.m. if you're watching in Sydney 8:40 a.m.. If you're catching us from Hong Kong, you can also tune in to the Bloomberg Australia podcast, delving into the bigger stories shaping the country's role in global business. Find that on Apple, Spotify, or Bloomberg.com. More ahead here. On the Asia trade. This is Bloomberg.